Factory steam production and heating and ventilation coupling control method based on neural network

Abstract

The invention relates to the technical field of industrial control, and discloses a factory steam production and heating and ventilation coupling control method based on a neural network, and the method comprises the steps: collecting the parameters of a source side, a load side and a state side of a multi-energy coupling gradient utilization system, and constructing a time sequence feature vector; then, a hard parameter sharing neural network model is utilized to decouple and output steam flow, heating and ventilation heat load and stratum recharge pressure response predicted values; according to the recharge pressure response and the stratum seepage physical model, the maximum recharge flow is calculated in real time to serve as a safety boundary; constructing a discrete state space model, and mapping a demand prediction value into a state variable reference trajectory; and finally, under the constraint of the maximum recharge flow, driving mechanisms such as a submersible pump and a peak shaving device to act through a control parameter sequence with the operation cost and tracking error minimization as targets by the discrete state space model. Geothermal exploitation safety and gradient cooperative utilization efficiency of different grades of heat energy are effectively considered.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to a neural network-based method for coupled control of steam and HVAC systems in factory production. Background Technology

[0002] Deep geothermal energy, as a clean and stable renewable energy source, has broad application prospects in the cascade utilization of industrial steam supply and building heating. A typical geothermal cascade utilization system usually first uses high-temperature geothermal fluids to produce industrial steam, then uses the heat-exchanged tailwater for building heating, and finally reinjects it into the underground reservoir. However, such systems involve the coupling of multiple physical domains, including deep geological environments, surface thermal equipment, and end-user loads, resulting in complex operating mechanisms and posing numerous challenges to practical control.

[0003] In existing control methods for multi-energy coupled systems, source-side and load-side predictions are typically modeled independently. This approach often overlooks the strong nonlinear coupling between geothermal extraction, industrial production scheduling, and building thermal inertia. For example, industrial steam demand is usually a rigid, rapidly changing pulse load, while building HVAC loads exhibit greater flexibility and hysteresis. Traditional methods struggle to effectively handle these two different time scales within a unified model framework, resulting in predictions that fail to accurately reflect the system's true dynamic response under various operating conditions, thus affecting the foresight and accuracy of the control strategy.

[0004] Furthermore, existing control strategies have significant shortcomings in terms of the safety of geothermal resource extraction. Conventional methods often use fixed thresholds determined during the design phase to limit extraction flow or reinjection pressure. However, the physical properties of deep geothermal reservoirs (such as permeability and skin factor) are not static but dynamically shift with factors such as changes in reinjection fluid temperature, blockage by suspended solids, or chemical precipitation. If control is based solely on static parameters, it can easily lead to geological safety risks such as wellhead overflow or bottomhole pressure exceeding rock fracturing pressure when reinjection resistance increases; conversely, overly conservative settings will limit the system's productivity and reduce energy utilization.

[0005] At the level of collaborative optimization control, current control systems mostly focus on meeting the setpoint tracking requirements of a single loop, lacking a global perspective on the entire cascade utilization chain. Existing control logic often struggles to dynamically adjust control weights according to different operating scenarios, failing to effectively balance the stability requirements of industrial steam supply with the energy-saving and consumption-reducing needs of HVAC systems. Especially when heat sources are insufficient or load fluctuations are large, the lack of an effective priority scheduling mechanism results in high-grade heat energy failing to prioritize critical industrial production, or failing to fully utilize the thermal energy storage characteristics of the building envelope to absorb system fluctuations, leading to high system operating costs and unstable energy supply quality. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a neural network-based method for coupled control of steam and HVAC systems in factory production. This method solves the problem that it is difficult to effectively handle the characteristics of these two different time scales within a unified model framework, resulting in prediction results that cannot accurately reflect the true dynamic response of the system under multiple operating conditions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a neural network-based method for the coupled control of steam production and HVAC in a factory. This method is based on a multi-energy coupled cascade utilization system. The multi-energy coupled cascade utilization system includes a physically connected deep geothermal energy supply loop, an auxiliary steam production loop containing a steam generator, and an HVAC loop. The method mainly includes: collecting source-side parameters, load-side parameters, and state-side parameters of the system; performing time alignment and standardization on the collected data to construct a time-series feature vector reflecting the current operating state of the system; inputting the time-series feature vector into a pre-trained neural network model to output predicted values ​​of steam flow demand, HVAC heat load demand, and formation reinjection pressure response within a preset future time period; calculating the maximum reinjection flow of the deep geothermal system at the current moment based on the predicted formation reinjection pressure response and a formation seepage physical model constructed based on Darcy's law and its non-Darcy flow correction theory; and determining the parameters used to characterize the auxiliary steam production loop. The system identifies the state variables of the operating state of the heating, ventilation, and air conditioning (HVAC) circuit and establishes a discrete state-space model describing the dynamic changes of these state variables. Using this discrete state-space model, the predicted steam flow demand and HVAC heat load demand are mapped to reference trajectories of the state variables. Under the constraint of maximum reinjection flow, and with the objectives of minimizing the operating cost of the multi-energy coupled cascade utilization system and minimizing the tracking error of the reference trajectory, the optimal control parameter sequence within a finite future time domain is solved. Based on this optimal control parameter sequence, the system controls the actuators, which include geothermal well submersible pumps, regulating valves, gas peak-shaving devices, and shallow geothermal heat pump units.

[0008] Preferably, the source-side parameters include the wellhead fluid temperature of the deep production well, the current production flow rate of the deep production well, the wellhead pressure of the deep reinjection well, and the liquid level of the deep reinjection well; the load-side parameters include the plant production scheduling time series vector, the real-time pressure of the steam network, the real-time flow rate of the steam network, the indoor average temperature, the outdoor dry-bulb temperature, and the solar radiation intensity; the status-side parameters include the opening feedback of each regulating valve, the operating frequency feedback of the submersible pump, the real-time operating power of the gas peak shaving device, and the start-up and shutdown status of the shallow geothermal heat pump unit.

[0009] In one specific embodiment, the neural network model is constructed using a hard parameter sharing mechanism. The model includes a shared feature extraction layer and a task-specific decoding layer. The shared feature extraction layer consists of multiple long short-term memory (LSTM) network units, used to extract shared latent space feature vectors from the time-series feature vectors. The task-specific decoding layer includes a steam rigid demand prediction branch, a HVAC flexible demand prediction branch, and a reinjection pressure response prediction branch. Each branch receives the shared latent space feature vector and independently outputs predicted steam flow demand, predicted HVAC heat load demand, and predicted formation reinjection pressure response.

[0010] Preferably, the process for determining the maximum reinjection flow rate includes: calculating the bottom-hole flowing pressure of the deep reinjection well based on the wellhead pressure in the source-side parameters and the hydrostatic equation of the wellbore; then, using an online identification algorithm, calculating the dynamic reinjection resistance coefficient based on historical bottom-hole flowing pressure data and historical production flow rate data; determining the maximum safe bottom-hole pressure threshold based on the formation rock fracture pressure parameter and obtaining the known formation static pressure; calculating the difference between the maximum safe bottom-hole pressure threshold and the formation static pressure, multiplying the difference by a preset safety margin coefficient, and dividing by the dynamic reinjection resistance coefficient to obtain the boundary of the maximum reinjectable flow rate on the source side; finally, taking the smaller value between the boundary of the maximum reinjectable flow rate on the source side and the rated maximum flow rate of the geothermal well submersible pump 4 as the maximum reinjection flow rate.

[0011] Furthermore, the dynamic reinjection impedance coefficient is calculated using a recursive least squares (RLS) algorithm with a forgetting factor. Specific steps include: constructing a linear regression model; calculating the gain vector based on the covariance matrix and forgetting factor from the previous time step; and updating the current dynamic reinjection impedance coefficient using the observed values ​​and the prior prediction error calculated based on the estimated values ​​from the previous time step.

[0012] Preferably, the optimal control parameter sequence is obtained by solving a quadratic programming (QP) problem within a model predictive control framework. The objective function of the quadratic programming problem comprises three parts: an output tracking deviation term, representing the L2 norm deviation between the predicted value and the reference setpoint of the state variable; a control increment term, representing the rate of change of the control variable; and a relaxation factor penalty term, representing the degree of violation of the output soft constraint. In the objective function, the weighting coefficient corresponding to the internal pressure of the steam generator 5 is set to be greater than the weighting coefficient corresponding to the heating and water supply temperature, to achieve a control strategy prioritizing steam supply stability.

[0013] Preferably, the constraints for solving the quadratic programming problem include: control variable amplitude constraints, limiting the real-time flow rate of the geothermal fluid to be less than or equal to the maximum reinjection flow rate; control variable increment constraints, limiting the rate of change of the real-time flow rate of the geothermal fluid to be within the range of physical acceleration and deceleration allowed by the actuator; and output soft constraints, which, by introducing a relaxation factor, allow the internal pressure of the steam generator 5 and the heating water supply temperature, which are state variables, to have slight fluctuations outside the preset hard constraint boundaries.

[0014] Preferably, the method of the present invention further includes the step of automatically adjusting the weight coefficients in the objective function according to the operating mode: when the system is in the heating and steam co-operation mode, the weight coefficient corresponding to the heating and water supply temperature in the objective function is set to a non-zero positive value, so that the controller simultaneously tracks and optimizes the internal pressure of the steam generator 5 and the heating and water supply temperature; when the system is in the pure industrial steam mode, the weight coefficient corresponding to the heating and water supply temperature in the objective function is set to zero, so that the controller only tracks and optimizes the internal pressure of the steam generator 5 and ignores the fluctuation of the heating and water supply temperature.

[0015] A second aspect of this invention provides a neural network-based factory production steam and HVAC coupled control system, which is applied to a neural network-based factory production steam and HVAC coupled control method. The system includes: a data acquisition and processing module for acquiring source-side, load-side, and state-side parameters of a multi-energy coupled cascade utilization system and constructing a time-series feature vector; a multi-task prediction module for running a neural network model and outputting predicted values ​​of steam flow demand, HVAC heat load demand, and formation reinjection pressure response based on the time-series feature vector; a dynamic boundary calculation module for calculating the maximum reinjection flow based on the predicted reinjection pressure response and a formation seepage physical model; a collaborative optimization module for mapping the demand prediction values ​​to a reference trajectory of state variables using a discrete state-space model and solving for the optimal control parameter sequence within a finite future time domain under the constraint of the maximum reinjection flow; and an execution control module for receiving the optimal control parameter sequence and driving the actuators, including the geothermal well submersible pump, regulating valve, gas peak-shaving device, and shallow geothermal heat pump unit, to operate.

[0016] This invention provides a neural network-based method for coupled control of steam and HVAC systems in factory production. It offers the following advantages: 1. This invention achieves decoupled sensing of multi-dimensional states on the source-load side by employing a neural network model based on a hard parameter sharing mechanism for multi-task prediction. It utilizes a shared Long Short-Term Memory (LSTM) network layer to extract common features from time-series data, and outputs steam flow demand, HVAC load demand, and formation reinjection pressure response through task-specific decoding layers. This structure effectively handles nonlinear correlations between different time scales and physical domains in multi-energy coupled systems, avoids feature conflicts between independent prediction models, and provides high-precision look-ahead information that balances production demand and formation response for subsequent optimized control.

[0017] 2. This invention ensures the operational safety and lifespan of deep geothermal systems by introducing a dynamic boundary calculation mechanism based on a physical model and online identification. Combining Darcy's law and non-Darcy flow correction theory, a recursive least squares method with a forgetting factor is used to update the dynamic reinjection impedance coefficient in real time. This allows for the calculation of the maximum reinjection flow rate at the current moment, taking into account the time-varying characteristics of formation properties. This flow rate value is used as a hard constraint to directly limit the amplitude of the control quantity, effectively preventing geological risks such as bottom hole pressure exceeding limits, formation fracturing, or reinjection overflow caused by pursuing instantaneous energy efficiency. This achieves proactive defense of the physical safety boundary in geothermal resource development.

[0018] 3. This invention utilizes a discrete state-space model combined with a quadratic programming solver to achieve multi-objective collaborative optimization control of industrial steam and building HVAC. By establishing a mathematical model describing the dynamic relationship between the internal pressure of the steam generator and the heating and water supply temperatures, and automatically adjusting the weight coefficients in the objective function according to the operating mode, priority response to rigid steam demands is ensured. This control strategy utilizes the thermal inertia of the HVAC system to absorb fluctuations, achieving a balance between minimizing system operating costs and minimizing the tracking error of each loop reference trajectory while satisfying the maximum reinjection flow constraint. This solves the problem that traditional control methods struggle to simultaneously consider the cascade utilization efficiency of different grades of thermal energy. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of the multi-energy coupling cascade utilization physical energy subsystem of the present invention; Figure 2 This is a schematic diagram of the process of the neural network-based factory production steam and HVAC coupling control method of the present invention; Figure 3 This is a schematic diagram of the overall architecture of the multi-energy coupling cascade utilization control system of the present invention.

[0020] The components include: 1. Deep mining well; 2. Primary plate heat exchanger; 3. Deep reinjection well; 4. Submersible pump; 5. Steam generator; 6. Gas peak shaving device; 7. Shallow geothermal heat pump unit; 8. Terminal air conditioning equipment; 9. Shallow geothermal heat exchanger pipe group; 10. Three-way regulating valve; 101. Data acquisition and processing module; 102. Multi-task prediction module; 103. Dynamic boundary calculation module; 104. Collaborative optimization module; and 105. Execution control module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example: Please see the appendix Figure 1 -Appendix Figure 2 This invention provides a neural network-based method for coupled control of steam and HVAC in factory production. This method is based on a multi-energy coupled cascade utilization system.

[0023] Please see the appendix Figure 1 The multi-energy coupled cascade utilization system includes a physical energy subsystem and a control system. The physical energy subsystem mainly consists of a deep geothermal energy supply circuit, an auxiliary steam generation circuit, and a heating, ventilation, and air conditioning (HVAC) circuit.

[0024] The deep geothermal energy supply loop includes a deep production well 1, a primary plate heat exchanger 2, a deep reinjection well 3, and a submersible pump 4. The submersible pump 4 is located inside or at the wellhead of the deep production well 1, and its outlet is connected to one side inlet of the primary plate heat exchanger 2 via a pipeline. One side outlet of the primary plate heat exchanger 2 is connected to the deep reinjection well 3 via a pipeline. Deep geothermal fluid is extracted by the submersible pump 4, flows through the primary plate heat exchanger 2 to release heat, and is then reinjected into the formation as tailwater.

[0025] The auxiliary steam generation circuit includes a steam generator 5 and a gas peak-shaving device 6. The secondary side outlet of the primary plate heat exchanger 2 is connected to the water inlet of the steam generator 5. The steam generator 5 is equipped with the gas peak-shaving device 6, which is used to supplement heat when the preheated feedwater heat is insufficient to meet process requirements.

[0026] The HVAC circuit includes a shallow geothermal heat pump unit 7, terminal air conditioning equipment 8, and a shallow geothermal heat exchange pipe group 9. The primary side outlet pipe of the first-stage plate heat exchanger 2 is connected to the water supply side of the terminal air conditioning equipment 8 through a three-way regulating valve 10, so that the deep geothermal tailwater after the first-stage heat exchange can directly enter the terminal air conditioning equipment 8 for heating, realizing the cascade utilization of energy.

[0027] The shallow geothermal heat pump unit 7 is connected to the supply and return water pipes of the terminal air conditioning equipment 8, and is also connected to the shallow geothermal heat exchange pipe group 9 buried underground. When the heat from the deep geothermal tailwater is insufficient to meet the needs of the terminal air conditioning equipment 8, the shallow geothermal heat pump unit 7 extracts heat energy from the shallow soil to supplement the heat.

[0028] Please see the appendix Figure 1 The control method of the present invention is executed by a control system, which is communicatively connected to various sensors and actuators in the aforementioned physical energy subsystem. The control system includes a data acquisition and processing module 101, a multi-task prediction module 102, a dynamic boundary calculation module 103, a collaborative optimization module 104, and an execution control module 105.

[0029] The data acquisition and processing module 101 connects to the factory's distributed control system and manufacturing execution system. The data acquired by the module includes the wellhead temperature and real-time flow rate of the deep production well 1, the wellhead pressure and liquid level of the deep reinjection well 3, the factory production schedule, the real-time pressure and flow rate of the steam pipeline network, the indoor temperature of the area where the terminal air conditioning equipment 8 is located, and outdoor environmental meteorological parameters. The module 101 also performs time alignment, noise reduction, cleaning, and vectorization processing on the acquired raw data to generate a time-series feature vector reflecting the current operating status of the system.

[0030] The multi-task prediction module 102 is connected to the data acquisition and processing module 101. The multi-task prediction module 102 internally stores and runs a neural network model built based on a deep learning algorithm. The multi-task prediction module 102 receives time-series feature vectors and outputs a multi-dimensional prediction sequence for a preset future time period. The multi-dimensional prediction sequence includes predicted flow demand values ​​for steam produced by the plant, predicted HVAC heat load demand values ​​required to maintain a set indoor temperature, and predicted formation reinjection pressure response values ​​under different reinjection flow rates.

[0031] The dynamic boundary calculation module 103 is connected to the multi-task prediction module 102. The dynamic boundary calculation module 103 calculates the maximum allowable reinjection flow rate under the current geological conditions based on the predicted formation reinjection pressure response and the formation seepage physical model. The maximum reinjection flow rate characterizes the physical limitations of the geological environment on system operation and serves as a safety boundary constraint for subsequent control optimization.

[0032] The collaborative optimization module 104 is connected to the dynamic boundary calculation module 103 and the multi-task prediction module 102. The collaborative optimization module 104 constructs a state-space equation describing the energy coupling relationship between the auxiliary steam generation loop and the HVAC loop. Under the constraints of the safety boundary of the maximum reinjection flow rate and the rigid demand for production steam, the collaborative optimization module 104 uses the minimization of the total operating cost of the control system as the objective function and employs a rolling time-domain optimization algorithm to solve for the optimal control parameter sequence of each actuator in the future finite time domain.

[0033] The execution control module 105 is connected to the collaborative optimization module 104. The execution control module 105 receives the optimal control parameter sequence and decomposes it into frequency conversion control commands for the submersible pump 4, opening control commands for each regulating valve, power control commands for the gas peak shaving device 6, and start / stop regulation commands for the shallow geothermal heat pump unit 7. The execution control module 105 sends these commands to the corresponding underlying controllers, driving the physical equipment to operate and achieving coupled control of the plant's production steam and HVAC systems.

[0034] Please see the appendix Figure 2 The method includes the following steps: S1. Construct a multi-source heterogeneous data acquisition and feature space. Periodically collect source-side parameters, load-side parameters, and state-side parameters of the energy system in the factory park to construct a time-series feature vector reflecting the current operating state of the system.

[0035] S2 performs multi-task prediction. The time-series feature vector is input into the neural network model, which outputs the predicted values ​​of steam flow demand, HVAC heat load demand, and formation reinjection pressure response for a preset time period in the future.

[0036] S3. Construct the dynamic flow boundary on the source side. Using the predicted value of the formation reinjection pressure response and combining it with the formation seepage physical model, calculate the maximum reinjection flow of the deep geothermal system at the current moment.

[0037] S4, Update the system state-space model. Update the discrete state-space model describing the energy transfer relationship between the steam generation side and the HVAC side based on real-time measurements.

[0038] S5 performs rolling time-domain optimization. Under the constraints of maximum reinjection flow rate and rigid demand for production steam, the optimal control sequence is solved using a discrete state-space model with the objective of minimizing system operating costs.

[0039] S6, Generate strategy and closed-loop control. The optimal control sequence is sent to the actuator, and steps S1 to S6 are repeated in the next control cycle based on the execution deviation.

[0040] In step S1, the process of constructing a multi-source heterogeneous data acquisition and feature space mainly includes a comprehensive digital mapping of the physical state, production plan, and environmental parameters of the factory park's energy system. This process is specifically divided into the following sub-steps: Step S101: Acquire source-side characteristic data. The system acquires real-time operating parameters of the deep geothermal system through a sensor network deployed at the geothermal wellhead. Specific data acquisition targets include the wellhead fluid temperature of deep production well 1. Current production flow rate of deep production well 1 The wellhead pressure of deep reinjection well 3 and the fluid level in deep reinjection wells The wellhead fluid temperature of deep production well 1 and the current production flow rate of deep production well 3 are used to calculate the total high-grade geothermal energy power available to the system at the current moment. The wellhead pressure and liquid level of deep reinjection wells are key indicators for assessing the dynamic changes in formation permeability and reinjection capacity. The acquisition of the above temperature, pressure, and flow rate data is achieved using resistance temperature sensors, pressure transmitters, and electromagnetic flowmeters. The signal transmission and analog-to-digital conversion technologies are well-known in the field and will not be elaborated here.

[0041] Step S102: Collect load-side characteristic data. Load-side data is divided into rigid production load data and flexible environmental load data. For rigid production loads, the system reads the scheduling plan from the Manufacturing Execution System (MES) and the steam network status from the Distributed Control System (DCS) via a communication interface. Since scheduling plans are typically discrete events or shift information, the system maps discrete production instructions to corresponding numerical load levels, or uses one-hot encoding to convert them into a production scheduling time series vector aligned with the control cycle. In conjunction with the real-time pressure of the steam pipeline network Real-time flow rate of steam pipeline network This constitutes the characteristics of the production load. For flexible environmental loads, the system collects the average indoor temperature of the area where the 8 terminal air conditioning units are located. Outdoor dry bulb temperature and solar radiation intensity These parameters are used to subsequently calculate the building's heat load requirements and thermal inertia potential.

[0042] Step S103: Collect state-side characteristic data. The system reads the feedback status of each actuator at the current moment to determine the initial state point of the control algorithm. The data collected includes the opening feedback of the regulating valves of each branch of the primary plate heat exchanger. Operating frequency feedback of submersible pumps and circulating pumps Real-time operating power of gas peak shaving device And the start-up and shutdown status of shallow geothermal heat pump units.

[0043] Step S104: Data Alignment and Feature Space Construction. Due to the different sampling frequencies of the data, the system downsamples the high-frequency data and interpolates or holds the low-frequency data to unify them to a preset control time step. After completing time alignment and standardization, the source-side, load-side, and state-side data are combined to construct... The system's global feature vector at time 1 .

[0044] System global feature vector Defined mathematically as:

[0045] In the formula: The source-side feature vector contains ; The load-side feature vector contains ; The state-side feature vector contains The system's global feature vector It includes information on the physical boundaries, demand trends, and equipment status of the system's historical operations, providing a complete input data foundation for multi-task prediction in subsequent steps. After completing the raw data acquisition in steps S101 to S103, given the significant differences in sampling frequency, signal quality, and dimensions among the source-side, load-side, and state-side data, step S104 further performs data preprocessing and multi-scale time alignment on the aforementioned multi-source heterogeneous data in order to construct a standardized neural network input interface. Step S104 specifically includes the following sub-steps: Step S1041 involves data cleaning and outlier handling. To address potential noise interference or data loss during sensor signal transmission, the completeness and validity of the acquired raw time-series data are verified. For detected missing data points, linear interpolation of adjacent valid data is used to fill in the gaps. Outlier data points exceeding the limits of physical equipment or statistical thresholds are identified as abnormal noise and removed. The outlier is then replaced with the mean of a sliding window to ensure the continuity and smoothness of the input data. The specific implementation of the sliding window filtering algorithm can be determined by those skilled in the art based on the actual signal-to-noise ratio; this is well-known in the field and will not be elaborated upon here.

[0046] Step S1042: Perform multi-scale time series alignment. Set the unified control period of the system. (For example, 5 to 15 minutes), the sampling frequency of all heterogeneous data is uniformly mapped to this reference time axis. This applies to the wellhead pressure of deep reinjection wells with sampling frequencies higher than the unified control period. Real-time pressure of steam pipeline network and real-time flow rate of steam pipeline network Downsampling is performed using the arithmetic mean method. Specifically, the calculation is performed for each unified control cycle. The arithmetic mean of all high-frequency sampling points within the time window is used as the effective feature value at the current control moment, thereby filtering out high-frequency fluctuation noise while preserving the signal trend characteristics. This is applied to production scheduling time series vectors with sampling frequencies lower than the unified control cycle. A zero-order hold method is employed. Specifically, within the same production shift or planned time period, the discrete production instruction value is kept unchanged until the next scheduling change, thereby expanding the step-like discrete planning data into a continuous time series synchronized with the unified control cycle. This is applied to outdoor dry-bulb temperatures with a sampling frequency lower than the unified control cycle but exhibiting a continuous gradual trend. and solar radiation intensity The linear interpolation method is used to calculate the values ​​at each moment on the reference time axis to fill the data gaps between adjacent low-frequency sampling points.

[0047] Step S143: Perform feature standardization. To eliminate the impact of differences in the dimensions and orders of magnitude of different physical parameters on the weight updates of the neural network model, a standardization transformation is performed on the time-aligned data for each dimension. The Z-score standardization method is used for any feature component. Its standardized value Calculate using the following formula: In the formula: This is the arithmetic mean of the feature component in the historical training dataset; This represents the standard deviation of the feature component in the historical training dataset. After the above cleaning, alignment, and standardization processes, the data is finally concatenated along its dimensions to generate... The system's global feature vector at time 1 .

[0048] After aligning and standardizing the data through the above steps, in order for the neural network model to capture the dynamic time-varying characteristics and long-term dependency patterns of the system, the system needs to construct a time series feature matrix based on a sliding time window. This process is achieved through the following step S105: Step S105, construct the sliding window time series feature matrix. Since the response of geothermal reinjection pressure and the changes in building indoor temperature both have time lag characteristics, the feature vector at a single moment cannot completely describe the dynamic state of the system. Therefore, the system sets the historical time window length. The length of this historical time window The numerical settings need to cover the system's critical response cycles, for example, set to 12 to 24 time steps to cover the past 1 to 2 hours of operation history.

[0049] Based on unified control cycle The system extracts the current moment from the historical database at discrete time points. and the past The system's global feature vectors at each time point are stacked in chronological order to form... Model input feature matrix at time step The mathematical expression of this matrix is ​​as follows: In the formula: For dimension A two-dimensional matrix; For the system's global feature vector The total number of feature dimensions is equal to the sum of the dimensions of the source-side feature vector, the load-side feature vector, and the state-side feature vector; This represents the length of the historical time window. System global feature vector. (in ) is the feature vector after standardization for the corresponding historical moment.

[0050] In the model's input feature matrix, different feature subspaces characterize physical laws in different dimensions: the time-series data of the source-side feature vectors characterize the historical seepage evolution trend of geothermal energy under geological conditions, supporting the model's learning of the nonlinear flow resistance characteristics of the strata; the time-series data of the load-side feature vectors characterize the fluctuation patterns of factory production loads and the gradient changes in environmental meteorology, supporting the model's identification of peak and valley patterns in the production cycle and prediction of building heat load demand; the time-series data of the state-side feature vectors characterize the historical motion trajectories of actuators, providing the model with the initial inertial state of the control system. The final constructed model input feature matrix... This data will be used as input to the deep neural network in the subsequent multi-task prediction steps, realizing the mapping from discrete sampling points in physical space to high-dimensional feature space.

[0051] The multi-task prediction model construction process involving geological constraints in step S2 is implemented using a deep neural network architecture with a hard parameter sharing mechanism. This neural network architecture is specifically constructed and implemented through the following sub-steps: Step S201: Construct a shared feature extraction layer. This is to extract the feature matrix from the model input. The coupling features between source-side, load-side, and state-side data are extracted, and a shared feature extraction layer is set in the model. The shared feature extraction layer consists of multiple stacked Long Short-Term Memory (LSTM) network units. The shared feature extraction layer receives the model input feature matrix. As input, the LSTM unit processes the feature vector at each time step sequentially along the time axis, updating its internal state through forget gates, input gates, and output gates. After processing the first... After processing data at each time step, the system extracts the hidden state vector of the last LSTM unit as the shared latent space feature vector. Shared latent space feature vectors The data numerically characterize the historical seepage trend of geothermal wells, the periodicity of factory production load, and the current thermal inertia state of buildings.

[0052] Step S202: Construct the task-specific decoding layer. After the shared feature extraction layer, the network structure branches into three independent task branches: a steam rigid demand prediction branch, a HVAC flexible demand prediction branch, and a reinjection pressure response prediction branch. Each task branch consists of multi-layer fully connected neural networks, with the number of neurons in each layer decreasing layer by layer along the signal propagation direction. Layers are connected by non-linear activation functions (such as ReLU or Swish functions). The steam rigid demand prediction branch maps the shared latent space feature vector to the steam flow trend required for the production process; the HVAC flexible demand prediction branch maps the shared latent space feature vector to the terminal heat load required to maintain the set room temperature; and the reinjection pressure response prediction branch maps the shared latent space feature vector to the pressure evolution trend of the formation under the current reinjection state.

[0053] Step S203: Define the network output vector. The output layers of all three task branches use linear activation functions to output continuous regression predictions. To support subsequent rolling time-domain optimization, the model is optimized for future prediction time domains. Each discrete time step within (in ), calculate the corresponding predicted values ​​respectively, and combine them to generate Multidimensional prediction output vector at time step The multidimensional prediction output vector The mathematical expression for it is defined as: ; In the formula: For the future Forecast values ​​of steam flow demand for factory production at each forecast time; For the future Forecast values ​​of HVAC heat load demand at each forecast time; For the future Predicted values ​​of formation reinjection pressure response at each predicted time.

[0054] Through the above architecture, the neural network model establishes a nonlinear mapping relationship from historical observation data to future multiphysics states. The weight matrix and bias parameters within the network are determined through backpropagation training using historical datasets. This network architecture utilizes a shared feature extraction layer to capture the cross-coupling effects of changes in geothermal extraction (source side) on reinjection pressure (geological side) and heating capacity (load side). The specific layer and neuron configurations in the neural network can be set by those skilled in the art based on the actual data scale and computational resource limitations; this is a conventional technique in the field and will not be elaborated upon here.

[0055] Step S3 involves the source-side dynamic boundary construction process based on reinjection capacity, which transforms the physical seepage characteristics of the geological layer into flow constraint boundaries executable by the control system. This process, based on data-driven prediction, introduces explicit hydrogeophysical equations as constraints, and is implemented through the following sub-steps: Step S301: Establish a geothermal radial flow physical model. To quantify the physical response mechanism between reinjection flow rate and bottom hole pressure, the system constructs a porous media radial steady-state flow model based on Darcy's Law and non-Darcy flow correction theory. In this model, the formation is considered a homogeneous and isotropic porous medium, and the functional relationship between reinjection flow rate and pressure difference is defined by wellbore hydrostatic formulas: ; In the formula: This is the bottomhole flowing pressure, a parameter based on the wellhead pressure of the deep reinjection well collected in step S1. The static pressure generated by the fluid column inside the wellbore is calculated after deducting friction loss along the way. This is the static pressure of the formation. This parameter is a known constant measured after the geothermal well is shut in and allowed to settle into equilibrium. It represents the original formation pressure outside the reinjection influence radius. This refers to the real-time flow rate of the geothermal fluid (i.e., the aforementioned instantaneous reinjection flow rate and extraction flow rate, which are equal in a closed system). The dynamic viscosity of the geothermal fluid; This is the formation volume factor; Formation permeability; The effective thickness of the aquifer; The coefficient is the non-Darcy turbulence coefficient.

[0056] Step S302: Identify the dynamic reinjection impedance coefficient. Considering that geological parameters are difficult to measure directly in real time, the system linearizes and simplifies the above physical formula near the current working point, and aggregates the relevant geological attribute parameters into a time-varying dynamic reinjection impedance coefficient. The simplified engineering calculation model is as follows: In the formula: This comprehensively reflects the equivalent fluid flow resistance characteristics under current geological conditions. The system calculates this coefficient in real time through the following sub-steps: Step S3021: Initialize the parameter identification algorithm and construct the regression vector. The system uses recursive least squares (RLS) with a forgetting factor to define the effective pressure difference observation value. Regression scalar with input as follows: Simultaneously initialize the covariance matrix. (In this embodiment, the parameter is a scalar in single-parameter identification), and its initial value is a large positive number. Step S322: Perform gain calculation and coefficient update. Process 1: Calculate the gain. : In the formula: It is the RLS forgetting factor.

[0057] Step 2: Update the dynamic reinjection impedance coefficient : ; Step 3: Update the covariance matrix : .

[0058] Step S303: Calculate the maximum safe reinjection flow rate boundary. To prevent the reinjection pressure from exceeding the formation's tolerance limit, the system sets a maximum safe pressure threshold at the bottom of the well. Based on real-time identification and safety margin coefficient (Value range 0.85 to 0.95), the system inversely solves for the maximum rechargeable flow boundary on the source side under the current geological conditions. : Note: The safety margin factor is mentioned here. Multiplication is performed on the numerator to ensure that the calculated upper limit of the flow rate is less than the theoretical limit, thus preserving a safety buffer. The calculated... This will serve as a hard constraint for subsequent optimization.

[0059] Step S4: Update the system state-space model. Establish a dynamic evolution model between the geothermal fluid input and the key thermodynamic state variables of the system. For the physical system consisting of the primary plate heat exchanger and steam generator in Step S1, this is implemented through the following sub-steps: Step S401: Construct a dynamic model for the steam generation stage. This stage utilizes high-temperature water from the secondary side of the primary plate heat exchanger to generate process steam within the steam generator. The system operates according to the law of conservation of energy, using the internal pressure of the steam generator... Differential equations are established for the state variables. The rate of change of steam pressure is mainly driven by the difference between the heat transfer input power (determined by geothermal flow) and the steam output power (determined by load). Its dynamic equation is described as follows: ; In the formula: The time derivative representing the internal pressure of a steam generator; The real-time flow rate of the geothermal fluid indirectly determines the heat input of the steam generator through the plate heat exchanger. The actual value of the steam flow demand for factory production is used as an external load disturbance. This is the system's self-balancing coefficient; This is the geothermal flow gain coefficient; This is the load disturbance factor.

[0060] Step S402: Construct a dynamic model for the HVAC (Heating, Ventilation, and Air Conditioning) stage. This stage utilizes the tailwater from the primary side outlet of the first-stage plate heat exchanger for direct heating. The system uses the heating supply water temperature... A model is established for the state variables. The rate of change of the supply water temperature depends on the residual enthalpy of the geothermal fluid and the return water state of the secondary network. The dynamic equation for this stage is described as follows: ; In the formula: The time derivative of heating water supply temperature; This refers to the return water temperature for heating. Heating circulating water flow rate The coefficient of thermal inertia attenuation; This represents the heat transfer gain coefficient.

[0061] Step S403: Generate discretized state-space equations. To adapt to the sampled control of a digital computer, the system combines the two continuous differential equations mentioned above and discretizes them using the zero-order hold method to generate a discrete state-space model for the MPC controller. Define the system state vector. Control input vector and external disturbance vector They are respectively: ; ; The complete discrete state-space model of the cascade utilization system is expressed as: ; ; In the formula: These are the state transition matrix, input control matrix, and disturbance response matrix obtained after discretizing the coefficient matrix of a continuous system, respectively. The system outputs an observation matrix (in this embodiment, an identity matrix). It uses operational data collected at preset time intervals to perform online correction of the matrix parameters via subspace identification.

[0062] Step S5 involves a collaborative scheduling process based on rolling time-domain optimization (MPC). This process generates optimal control commands by constructing and solving a quadratic programming (QP) problem that includes production process indicators, energy consumption, and equipment motion constraints. The specific implementation steps of this process are as follows: Step S501: Generate the dynamic setpoint trajectory. At each moment... Based on the future demand prediction sequence generated in step S2, the system reverse-calculates the reference trajectory of the state variables. For the steam stage, to ensure flash evaporation efficiency and meet the downstream minimum pressure requirements, the system sets a steam generator pressure reference trajectory. This trajectory is typically set to a fixed value higher than the minimum inlet pressure of the downstream pipeline network, or based on the actual value of the predicted plant production steam flow demand. Feedforward compensation adjustments are performed. For the heating component, the system uses the predicted HVAC heat load demand output in step S242. By utilizing the relationship between heat exchanger efficiency and the number of heat transfer units (NTU) or the inverse operation of the logarithmic mean temperature difference (LMTD) formula, the required reference trajectory for heating and water supply temperatures can be calculated. This trajectory represents the theoretical sequence of secondary water supply temperatures required to maintain a constant indoor set temperature within the future predicted time domain.

[0063] Step S502: Construct a multi-objective weighted optimization function. The system defines a scalar performance index function. This function is recalculated at each sampling time. It consists of an output tracking deviation term, a control increment term, and a soft constraint penalty term composed of relaxation factors. Its mathematical expression is as follows: ; In the formula: For prediction in the time domain, the value is typically taken as 20 to 50 sampling periods; To control the time domain, satisfy ; For the prediction of the first state based on the state-space model in step S4 The system output vector of the step includes elements such as the internal pressure of the steam generator. and heating and water supply temperature ; For the first The reference trajectory vector of the step, by composition; For the first The step-by-step control increment vector corresponds to the change in the geothermal wellhead extraction flow rate. ; The output error weight matrix is ​​a diagonal matrix. To ensure production safety, the weight element values ​​for the corresponding pressure are set to be greater than the weight element values ​​for the corresponding temperature (e.g., a ratio of 2 to 5), thereby establishing a control strategy that prioritizes steam supply stability over heating regulation. The control smoothing weight matrix is ​​used to limit drastic changes in control quantities and prevent mechanical wear caused by frequent adjustments of the reinjection pump and frequency converter. This is a relaxation factor used to soften the output constraints; To relax the penalty weights, take the maximum positive value (e.g., 10). 5 This ensures that violations of soft constraints are only permitted in extreme, unsolvable conditions.

[0064] Step S503, integrate constraints. The above scalar performance index function The minimization process is subject to the following constraints to ensure the physical feasibility and geological safety of the control commands: Control amplitude constraints: In the formula: This represents the wellhead extraction flow rate. Since this embodiment uses a closed-loop geothermal system, it is assumed that the wellhead extraction flow rate and the reinjection flow rate are approximately equal. Therefore, the comprehensive boundary of the reinjection side is directly used to limit the control variables of the extraction side, thereby ensuring that the extraction operation is always within the geological safety boundary and equipment capacity range.

[0065] Incremental constraints on control variables: In the formula: and These are the maximum allowable descent rate and maximum ascending rate of the actuator, respectively, determined by the acceleration and deceleration time parameters of the frequency converter.

[0066] Output soft constraints: In the formula: and These represent the lower and upper limits of the permissible state variables (pressure and temperature) for the process, respectively. A non-negative relaxation factor is introduced. The system transforms the original hard constraints into soft constraints, ensuring the solvability of the optimization problem under any perturbation.

[0067] Step S6 involves the strategy execution logic and automatic mode switching process. This process is responsible for converting the optimized control commands calculated in step S5 into physical actions of the underlying actuators, and automatically adjusting the control strategy structure according to changes in the external environment or system state. The system implements the above functions through a hierarchical control architecture and state machine logic, specifically through the following sub-steps: Step S601: Execute low-level flow following control. Considering that the actuator (frequency converter) of the reinjection pump is directly controlled by the power supply frequency, while the geothermal wellhead extraction flow command... To achieve precise command execution for the physical value of the flow rate, this embodiment employs a cascade control structure. The system uses the MPC controller as the outer loop and the local PLC (Programmable Logic Controller) as the inner loop. The data is transmitted via fieldbus to the local PLC as the setpoint (SP) for the flow PID control loop. The local PLC acquires the real-time feedback value (PV) of the electromagnetic flowmeter at microsecond intervals and uses a proportional-integral (PI) control algorithm to calculate the inverter's operating frequency command. This PI algorithm eliminates steady-state error through integral action, ensuring that the actual geothermal flow can quickly and accurately track the geothermal wellhead extraction flow command issued by the MPC, thereby overcoming the open-loop control error caused by the time-dependent drift of pipeline resistance characteristics.

[0068] Step S602: Implement multi-condition mode switching. Given the significant seasonal operating characteristics of the geothermal cascade utilization system, the system incorporates state machine logic to dynamically reconstruct the output error weight matrix in step S502 based on date signals or manually input mode commands. When the system is in "heating and steam co-operation mode" (usually in winter), the system setting matrix... The diagonal elements are non-zero positive values, for example... At this time, the MPC controller simultaneously monitors the internal pressure of the steam generator. and heating and water supply temperature Tracking and optimization are performed. When the system switches to "pure industrial steam mode" (usually in summer), the system automatically adjusts the matrix. The weighted elements corresponding to the heating and water supply temperatures are set to zero, i.e. And remove the soft constraint on temperature in the output. At this point, optimize the scalar performance index function. The temperature deviation term is no longer included. The MPC controller will ignore changes in the supply water temperature and focus only on the stable control of the internal pressure of the steam generator. This ensures steam supply while avoiding unnecessary power consumption of the reinjection pump due to adjustments in heating temperature. At the moment of mode switching, to prevent system oscillation caused by sudden changes in control parameters, the system uses linear interpolation to smoothly transition the weighted parameters from their current values ​​to the target values ​​within a preset time window (e.g., 10 minutes), achieving a seamless switching of the control strategy.

[0069] Step S603: Execute the anomaly monitoring and degradation strategy. To ensure system operational safety, the system monitors the status feedback of the MPC solver and the validity of sensor data in real time. If optimization fails (e.g., the solver cannot converge within the maximum number of iterations) or a critical sensor malfunctions (e.g., pressure or flow signal disconnection), the system will automatically trigger a fault-tolerant control strategy. In fault-tolerant mode, the system forcibly locks the MPC output, maintaining the previously valid geothermal wellhead production flow command. The system remains unchanged; or it automatically switches to a backup PID controller, which regulates solely to maintain the geothermal wellhead pressure within a safe range until the fault alarm is cleared or manual intervention is required. The specific code implementation of the aforementioned underlying PID control and state machine logic can be accomplished by those skilled in the art using existing standard modules of industrial automation configuration software, and is therefore well-known and will not be elaborated upon here.

Claims

1. A neural network-based method for controlling steam production and HVAC coupling in a factory, implemented using a multi-energy coupled cascade utilization system, wherein the multi-energy coupled cascade utilization system comprises a physically connected deep geothermal energy supply circuit, an auxiliary steam production circuit, and an HVAC circuit, characterized in that... Includes the following steps: The source-side parameters, load-side parameters, and state-side parameters of the multi-energy coupled cascade utilization system are collected, and the collected data are time-aligned and standardized to construct a time-series feature vector reflecting the current operating state of the multi-energy coupled cascade utilization system. The time-series feature vector is input into a pre-trained neural network model, which outputs the predicted values ​​of steam flow demand, HVAC heat load demand, and formation reinjection pressure response within a preset time period. Based on the predicted value of the formation reinjection pressure response, and combined with the formation seepage physical model constructed based on Darcy's law and its non-Darcy flow correction theory, the maximum reinjection flow rate of the deep geothermal system at the current moment is calculated as the safety boundary. Determine the state variables used to characterize the operating states of the auxiliary steam generation circuit and the HVAC circuit, and establish a discrete state-space model describing the dynamic relationship of the state variables; Using the maximum reinjection flow rate as a safety constraint condition for the dynamic flow boundary on the source side, and under the premise of meeting the rigid requirement of the predicted flow demand of steam for plant production, the discrete state space model is used to solve the optimal control parameter sequence of each actuator in the future finite time domain with the objective function of minimizing system operating cost. The optimal control parameter sequence is decomposed into control commands for the underlying controller and sent to the actuator to drive the physical device to achieve coupled control.

2. The neural network-based method for coupled control of steam and HVAC in factory production, as described in claim 1, is characterized in that... The deep geothermal energy supply circuit includes a deep mining well (1), in which a submersible pump (4) is installed. The outlet of the submersible pump (4) is connected to the inlet of one side of the primary plate heat exchanger (2) through a delivery pipe. The outlet of one side of the primary plate heat exchanger (2) is connected to the deep reinjection well (3) through a pipeline. The auxiliary steam generation circuit includes a steam generator (5) and a gas peak shaving device (6). The secondary outlet of the primary plate heat exchanger (2) is connected to the inlet of the steam generator (5). The HVAC circuit includes a shallow geothermal heat pump unit (7), terminal air conditioning equipment (8) and a shallow geothermal heat exchange pipe group (9). The primary outlet pipe of the primary plate heat exchanger (2) is connected to the water supply side of the terminal air conditioning equipment (8) through a three-way regulating valve (10).

3. The neural network-based method for coupled control of steam and HVAC in factory production, as described in claim 2, is characterized in that... The source-side parameters include the wellhead fluid temperature of the deep production well (1), the current production flow rate of the deep production well (1), the wellhead pressure of the deep reinjection well (3), and the liquid level of the deep reinjection well (3); the load-side parameters include the plant production scheduling time series vector, the real-time pressure of the steam network, the real-time flow rate of the steam network, the indoor average temperature, the outdoor dry-bulb temperature, and the solar radiation intensity; the status-side parameters include the opening feedback of the three-way regulating valve (10), the operating frequency feedback of the submersible pump (4), the real-time operating power of the gas peak shaving device (6), and the start-up and shutdown status of the shallow geothermal heat pump unit (7).

4. The neural network-based steam and HVAC coupling control method for factory production as described in claim 1, characterized in that, The neural network model is constructed using a hard parameter sharing mechanism, including a shared feature extraction layer and a task-specific decoding layer; The shared feature extraction layer consists of multiple long short-term memory network units and is used to extract the shared latent space feature vectors in the temporal feature vectors. The task-specific decoding layer includes a steam rigid demand prediction branch, a HVAC flexible demand prediction branch, and a reinjection pressure response prediction branch. Each branch receives the shared latent space feature vector and outputs the predicted steam flow demand, the predicted HVAC heat load demand, and the predicted formation reinjection pressure response.

5. The neural network-based method for coupled control of steam and HVAC in factory production, as described in claim 2, is characterized in that... The process of determining the state variables and constructing the discrete state-space model includes: The internal pressure of the steam generator (5) is selected as one of the state variables, and a dynamic equation for the steam generation stage is constructed, wherein the rate of change of the internal pressure of the steam generator (5) is related to the real-time flow rate of the geothermal fluid and the actual value of the steam flow rate required by the factory. The heating water supply temperature of the shallow geothermal heat pump unit (7) is selected as the second of the state variables, and a dynamic equation for the heating, ventilation and heating stage is constructed. The rate of change of the heating water supply temperature is related to the real-time flow rate of the geothermal fluid, the heating return water temperature and the heating circulating water flow rate. The dynamic equations of the steam generation stage and the dynamic equations of the heating, ventilation and air conditioning (HVAC) stage are combined and discretized to obtain the discrete state-space model containing the state transition matrix, the input control matrix, and the disturbance response matrix.

6. The neural network-based steam and HVAC coupling control method for factory production as described in claim 1, characterized in that, The process of achieving the maximum reinjection flow includes: Based on Darcy's law and the non-Darcy flow correction theory, a linearized engineering calculation model is established to describe the relationship between the flow pressure of deep reinjection wells (3) and the static pressure of the formation, the dynamic reinjection resistance coefficient and the real-time flow rate of geothermal fluid. The recursive least squares method with a forgetting factor is used to calculate the gain and update the estimated value of the dynamic reinjection impedance coefficient online by using the real-time collected effective differential pressure observation value and the input regression scalar. Set the maximum safe pressure threshold and safety margin coefficient for the deep reinjection well (3), and use the updated dynamic reinjection impedance coefficient to calculate the difference between the maximum safe pressure threshold of the deep reinjection well (3) and the static pressure of the formation, and perform inverse calculation in combination with the safety margin coefficient to obtain the current source side maximum reinjection flow boundary.

7. A neural network-based method for coupled control of steam and HVAC in factory production, as described in claim 1, is characterized in that... The constraints include: Control the amplitude constraint, directly use the maximum reinjection flow boundary to limit the deep mining well (1) mining flow, and ensure that the mining operation is always within the geological safety boundary; Control increment constraints are set based on the acceleration and deceleration time parameters of the frequency converter to determine the maximum allowable descent rate and maximum ascent rate of the actuator. Output soft constraints, set the lower and upper limits of the state variables allowed by the process, and combine them with non-negative relaxation factors to transform hard constraints into soft constraints.

8. The neural network-based method for coupled control of steam and HVAC in factory production according to claim 1, characterized in that, The objective function for minimizing the system operating cost consists of a multi-objective weighted term, specifically including: The output tracking deviation term is used to calculate the weighted sum of squares of the Euclidean distance between the system output vector and the reference trajectory vector. The weight element value corresponding to the pressure is set to be greater than the weight element value corresponding to the temperature, so as to establish a strategy that prioritizes the stability of steam supply. The control increment term is used to calculate the weighted sum of squares of the control increment vector to limit the range of motion of the actuator. The soft constraint penalty term, which consists of relaxation factors, allows for the softening of output constraints under extreme conditions by introducing relaxation penalty weights, thus ensuring that the optimization problem is solvable.

9. A neural network-based method for coupled control of steam and HVAC in factory production, as described in claim 2, is characterized in that... The method also includes automatically adjusting the weight coefficients in the objective function according to the operating mode: When the multi-energy coupled cascade utilization system is in the heating and steam co-operation mode, the weight coefficient corresponding to the heating and water supply temperature in the objective function is set to a non-zero positive value, so that the controller can simultaneously track and optimize the internal pressure of the steam generator (5) and the heating and water supply temperature. When the multi-energy coupled cascade utilization system is in pure industrial steam mode, the weight coefficient of the heating and water supply temperature in the objective function is set to zero, so that the controller only tracks and optimizes the internal pressure of the steam generator (5) and ignores the fluctuation of the heating and water supply temperature.

10. A neural network-based steam and HVAC coupled control system for factory production, characterized in that, The neural network-based plant production steam and HVAC coupling control method described in any one of claims 1-9 includes: The data acquisition and processing module (101) is connected to the factory's distributed control system and manufacturing execution system, and collects data from the source side, load side and status side and performs cleaning, alignment and vectorization processing. The multi-task prediction module (102) is connected to the data acquisition and processing module (101). It runs a neural network model with shared hard parameters and outputs predicted values ​​of steam flow demand, HVAC load demand and formation reinjection pressure response. The dynamic boundary calculation module (103) is connected to the multi-task prediction module (102) and is used to calculate the maximum recharge flow rate based on the predicted value of the formation recharge pressure response and the formation seepage physical model. The collaborative optimization module (104) is connected to the dynamic boundary calculation module (103) and the multi-task prediction module (102), and is used to solve the optimal control parameter sequence that minimizes the system operating cost based on the discrete state space model under the constraints of safety boundary and production demand. The execution control module (105) is connected to the collaborative optimization module (104) and is used to receive the optimal control parameter sequence and convert it into control commands to drive the submersible pump (4), the three-way regulating valve (10), the gas peak shaving device (6) and the shallow geothermal heat pump unit (7).

Images